Ask any AI tool about the humanitarian situation in Africa, and you will get all sorts of answers in seconds. Then check closely where those answers come from, and you will be left mesmerised.

That very question led me to examine 1,083 verifiable citations used by ChatGPT, Gemini, DeepSeek and Perplexity when answering humanitarian questions about ten African countries. The results revealed that only 5.5% of the citations came from organisations based in the country being discussed, while 68.8% came from United Nations agencies or other multilateral organisations. In other words, the crisis may be local, let’s say in Sudan or Kenya, but the institutions most visible when AI explains it are usually international.

For years, humanitarian localisation or simply put, the shifting of power and resources to local actors in affected countries, has focused on who receives funding, who implements programmes and who participates in decisions. The Grand Bargain, for example, committed major humanitarian actors to increasing support for local and national responders. Yet recent analysis by ODI Global shows that large gaps remain between localisation commitments and funding practice.

As AI becomes an alternative way people learn about crises, another dimension is worth examining: Whose knowledge gets seen?

A new layer of humanitarian power

The African Humanitarian AI Citation Index examined questions about Burkina Faso, Chad, the Democratic Republic of the Congo, Ethiopia, Kenya, Mozambique, Nigeria, Somalia, South Sudan and Sudan. The questions covered conflict, displacement, food insecurity, health emergencies, water and sanitation, climate shocks, humanitarian access, protection, livelihoods and humanitarian response.

My interest was not simply whether the AI answers were accurate. I wanted to see which organisations appeared as the sources behind those answers. I wanted to see which organisations appeared as the sources behind those answers. Those sources can shape where people look next and, eventually, which organisations they come to see as authoritative.

And, I wanted to see which organisations appeared as the sources behind those answers. In this study, visibility simply refers to which organisations were actually cited. That visibility was concentrated among a relatively small group of organisations.

The five most-cited knowledge producers accounted for 40.3% of all verifiable citations, while the top ten accounted for 58.8%. UNICEF was cited most often, followed by UNHCR, OCHA/Humanitarian Country Team products, OCHA and the International Organisation for Migration. Their prominence is understandable as these organisations publish extensively. In addition, they maintain established websites and archives, and have long-standing authority within the humanitarian system.

The more important question is why credible sources closer to the crises themselves remain so much harder to see.

Kenya: an interesting contrast

The country results were not uniform. Of the 60 country-local citations identified across the ten-country dataset, 31 came from Kenyan sources. Kenya alone accounted for more than half of all local citations found in the study. Nigeria also recorded relatively stronger local visibility.

The picture changed sharply elsewhere. Among 100 verifiable citations concerning the Democratic Republic of the Congo, I did not identify a single citation from an organisation based in the country. Sudan also recorded no country-local citation in its recoverable set, while Somalia recorded only one.

I cannot say from this study exactly why those differences occurred. Language is likely to play a part. So might publishing practices, search visibility, website authority, technical accessibility and the way organisations structure and attribute their information, as well as the fact that a large number of humanitarian organisations operate out of Nairobi as their regional office.

What the findings do show is that local visibility is not equally weak everywhere. That raises the next question: Why is locally produced knowledge easier for AI systems to find and cite in some African countries than in others?

Kenya provides a useful place to begin looking for answers.

Being online does not always mean being visible

African organisations already produce a great deal of knowledge through government data, local reporting, university research, civil society documentation and humanitarian field assessments. Some of this material contains knowledge that is difficult to reproduce from outside the communities concerned. Yet putting information online does not guarantee that it will become visible when an AI system builds an answer.

A local organisation may understand a community deeply and still have fewer resources to publish and distribute its evidence than a large international institution. This is part of a wider discoverability challenge I explored in an article on ‘How AI Search Is Changing Trust, Visibility, and Public Relations in Africa’ institutions can publish valuable information online and still struggle to become visible when AI systems decide which sources to surface.

The humanitarian information environment adds another layer to that challenge. International organisations often have dedicated information teams and strong distribution networks. They have been easier to find long before we ever thought of generative AI. AI operates within that existing information environment. When it repeatedly retrieves and cites organisations that are already highly visible online, existing patterns of authority can become even more pronounced.

The question is therefore broader than whether AI can find African sources. We also need to understand why some organisations have a much easier route into the digital knowledge systems from which AI draws.

What the findings tell us — and what they do not

The results of my study need to be read carefully. They do not show that AI systems deliberately discriminate against African or local sources. The study also cannot tell us how many suitable local sources were available for every question. All the questions were asked in English, which is an important limitation when studying countries where French, Portuguese, Arabic and other languages are widely used. Source-level citation capture was also incomplete for some of the AI platforms.

To check whether this affected the overall pattern, I examined ChatGPT separately because source-level information was available for all 90 production questions. In that complete dataset, 7.2% of citations came from sources based in the affected country, while 70.9% came from UN or multilateral organisations. The pattern was therefore similar.

The study makes a narrower claim as it shows which sources users were visibly presented with during these searches. That is worth paying attention to as AI becomes another gateway to humanitarian information.

Localisation may also be about whose knowledge travels

The whole idea of humanitarian localisation is all about moving power closer to the people and institutions most affected by crises. And as AI becomes part of the information environment, visibility may need to become part of that conversation too.

Local organisations need the capacity to publish evidence in ways that make it easier to find, verify and attribute. International organisations can help by linking clearly to the local research, data and field knowledge they use. AI companies also have reason to examine whether their systems repeatedly surface the same dominant information providers when credible local evidence exists.

The goal is not to create a quota for local citations. Humanitarian information still needs to be accurate, current and relevant. The issue is whether strong local evidence has a reasonable chance of being seen. Knowledge can exist and still struggle to travel.

Humanitarian localisation has rightly asked who delivers aid, who receives funding and who gets a voice in decisions. As AI becomes another starting point for understanding crises, one more question is worth adding: Who gets to explain African crises?